Pillowfort update:
So the current build of Pillowfort, v42, at least of last night. The one that's supposed to be going into beta. It works, but it breaks under ambition.
Tell the system you're a well funded type A, and want to start an eleven billion dollar oil company from scratch, and it starts to fray a little bit.
Makes sense. This is my fault. The mental model I designed with was `cute consulting company or research lab`.
Portia, the planner agent hallucinated an org chart with eleven hundred agents on it, and invented departments and department heads.
In practical terms... that would have meant a 13 hour standup every day.
It's only wrong because I did plan concretely for this kind of ambition.
Second experiment was to provide another scenario we hadn't planned for.
Same type A personality input. Same documentation.
Only this time, we're augmenting a department at an existing oil company.
Portia returned a much more sensible plan.
But the agent did something I didn't expect. It asked for a list of humans to enter into the org chart, along with their contact info.
The system was always theoretically capable of augmenting a human team. And it is something the agent should theoretically be able to do.
I just hadn't considered what would happen if we tried.
This tells me that the planner is working beautifully.
We're in the last mile, folks.